pith. sign in

arxiv: 2605.15556 · v1 · pith:GE2JY7QRnew · submitted 2026-05-15 · 💻 cs.HC

TopoClaw: A Human-Centric and Topology-Aware Agent Operating System

Heyuan Huang , Yeyi Guan , Jihong Wang , Mingzhi Wang , Jiamu Zhou , Xiangmou Qu , Jiaxin Yin , Xin Liao
show 2 more authors
Xingyu Lou Jun Wang
This is my paper

Pith reviewed 2026-05-20 17:12 UTC · model grok-4.3

classification 💻 cs.HC
keywords agentdevicedistributedtopoclawacrossactionsautonomouscontext
0
0 comments X

The pith

TopoClaw is a human-centric Agent OS that uses physical and social topology modeling to enable cross-boundary execution with identity attribution and context-aware governance.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Large language models are making AI assistants more capable at reasoning and completing complex tasks over time. This progress has created a need for Agent Operating Systems, which are like the underlying software that manages these AI agents, handling things like memory and deciding what tools to use. Most existing designs for these systems focus mainly on the AI agent itself, running on a single computer or device. TopoClaw is different because it puts humans at the center and considers the connections between things. It models the user's world in two ways: the physical setup of different devices and the social connections like teams or shared responsibilities. The system allows actions to be carried out on the best device for the job, even if the user is not directly using it. It also lets agents act as digital stand-ins for people in group settings, keeping track of who is doing what and respecting permissions. Finally, it has rules that adjust based on the situation to make sure the AI doesn't overstep boundaries. This paper explains the ideas behind TopoClaw, how it works in practice, how it handles security, and how it might be used in the future. It's meant as a starting point for building such systems.

Core claim

TopoClaw unifies device operation, messaging, and skills around accountable cross-boundary execution, with three core contributions: (1) cross-device action placement, decoupling intent from actuation and routing distributed actions across the device cluster based on hardware affordances and user context; (2) cross-user identity attribution, treating agents as socially situated Digital Twins that coordinate in multi-user spaces while preserving provenance, role-aware permissions, and human accountability; (3) cross-context authority governance, pairing broad capability with distributed, context-aware policy enforcement across physical and social trust boundaries to bound proactive autonomy at the OS layer.

Load-bearing premise

That modeling the user's ecosystem as two coupled structures (physical device topology of heterogeneous surfaces and social relationship topology of shared spaces, teams, and delegated roles) is sufficient to enable unified accountable cross-boundary execution and to bound proactive autonomy at the OS layer. (Abstract, paragraph describing the three core contributions.)

Figures

Figures reproduced from arXiv: 2605.15556 by Heyuan Huang, Jiamu Zhou, Jiaxin Yin, Jihong Wang, Jun Wang, Mingzhi Wang, Xiangmou Qu, Xingyu Lou, Xin Liao, Yeyi Guan.

Figure 1
Figure 1. Figure 1: Paradigm shift from traditional agent-centric isolation to TopoClaw’s human-centric dual-topology [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: TopoClaw System Architecture. The OS acts as a decoupled runtime navigating dual topologies. Intent [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
read the original abstract

Large language models (LLMs) have evolved AI assistants into autonomous reasoning engines that maintain context, invoke tools, and pursue long-horizon tasks. This has spurred Agent Operating Systems (Agent OS) as kernel-like layers for lifecycle management, memory, scheduling, and access control. Yet most designs remain agent-centric, treating the OS as a single-host runtime for internal reasoning and tool use, leaving open how autonomous actions integrate with distributed, collaborative, permission-sensitive workflows. TopoClaw is an open-source, human-centric, topology-aware Agent OS modeling the user's ecosystem as two coupled structures: a physical device topology of heterogeneous surfaces and a social relationship topology of shared spaces, teams, and delegated roles. It unifies device operation, messaging, and skills around accountable cross-boundary execution, with three core contributions: (1) cross-device action placement, decoupling intent from actuation and routing distributed actions across the device cluster based on hardware affordances and user context; (2) cross-user identity attribution, treating agents as socially situated "Digital Twins" that coordinate in multi-user spaces while preserving provenance, role-aware permissions, and human accountability; (3) cross-context authority governance, pairing broad capability with distributed, context-aware policy enforcement across physical and social trust boundaries to bound proactive autonomy at the OS layer. This report presents TopoClaw as an engineering-oriented reference architecture, covering its design principles, runtime, cross-device execution, collaboration mechanisms, security model, and deployment outlook.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes TopoClaw as an open-source, human-centric, topology-aware Agent Operating System. It models the user's ecosystem as two coupled structures—a physical device topology of heterogeneous surfaces and a social relationship topology of shared spaces, teams, and delegated roles—and unifies device operation, messaging, and skills around accountable cross-boundary execution. The three core contributions are: (1) cross-device action placement that decouples intent from actuation and routes distributed actions across the device cluster based on hardware affordances and user context; (2) cross-user identity attribution that treats agents as socially situated Digital Twins coordinating in multi-user spaces while preserving provenance, role-aware permissions, and human accountability; (3) cross-context authority governance that pairs broad capability with distributed, context-aware policy enforcement across physical and social trust boundaries to bound proactive autonomy at the OS layer. The report presents this as an engineering-oriented reference architecture covering design principles, runtime, cross-device execution, collaboration mechanisms, security model, and deployment outlook.

Significance. If the proposed reference architecture can be realized with concrete mechanisms and subsequently validated, it would provide a structured framework for integrating autonomous LLM-based agents into distributed, multi-device, and multi-user environments while emphasizing human accountability and bounding proactive behavior. This could advance HCI and distributed systems research by shifting from agent-centric runtimes to topology-aware designs that explicitly address cross-boundary execution, identity, and governance.

major comments (3)
  1. [Abstract] Abstract (paragraph describing the three core contributions): The central claim that modeling the ecosystem as two coupled physical-device and social-relationship topologies is sufficient to unify execution and bound proactive autonomy at the OS layer is asserted without any representation for the topologies, algorithm for cross-device action placement or routing, policy language, or conflict-resolution procedure for authority governance. This modeling assumption is load-bearing for all three contributions yet remains unelaborated.
  2. [cross-context authority governance] Section on cross-context authority governance: No analysis is provided of how provenance or permissions are maintained when topologies are incomplete or change, nor is there a concrete mechanism for distributed policy enforcement across trust boundaries. Without these, the claim to bound proactive autonomy reduces to an untested assertion.
  3. [cross-device action placement] Section on cross-device action placement: The description of routing distributed actions based on hardware affordances and user context lacks any formal topology representation, decision procedure, or handling for dynamic device clusters, which is required to support the decoupling of intent from actuation.
minor comments (2)
  1. [cross-user identity attribution] The term 'Digital Twins' is used without citation to prior literature on digital twins in HCI or multi-agent systems.
  2. [Design principles] Diagrams illustrating the two coupled topologies and their interaction with the runtime would substantially improve clarity of the reference architecture.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive feedback and recognition of TopoClaw's potential to advance topology-aware designs in Agent OS research. We address each major comment below, agreeing where elaboration is needed for a reference architecture and outlining specific revisions to strengthen the manuscript without altering its conceptual focus.

read point-by-point responses

  1. Referee: [Abstract] Abstract (paragraph describing the three core contributions): The central claim that modeling the ecosystem as two coupled physical-device and social-relationship topologies is sufficient to unify execution and bound proactive autonomy at the OS layer is asserted without any representation for the topologies, algorithm for cross-device action placement or routing, policy language, or conflict-resolution procedure for authority governance. This modeling assumption is load-bearing for all three contributions yet remains unelaborated.

    Authors: We agree that the abstract presents the modeling assumption at a high level without explicit representations or procedures. The manuscript frames TopoClaw as an engineering-oriented reference architecture focused on design principles rather than a complete algorithmic specification. To address this, we will revise the abstract and add a concise topology modeling subsection that defines the physical topology as an attributed graph of device surfaces and the social topology as a role-based relational structure. We will also sketch the action placement as a context-driven matching process and outline a basic policy language with hierarchical conflict resolution. revision: yes

  2. Referee: [cross-context authority governance] Section on cross-context authority governance: No analysis is provided of how provenance or permissions are maintained when topologies are incomplete or change, nor is there a concrete mechanism for distributed policy enforcement across trust boundaries. Without these, the claim to bound proactive autonomy reduces to an untested assertion.

    Authors: The referee correctly identifies that the governance section remains high-level and lacks analysis of dynamic or incomplete topologies. The current manuscript describes the security model conceptually but does not detail maintenance or enforcement mechanisms. We will revise by adding a subsection on dynamic topology handling, including provenance via immutable audit logs tied to Digital Twin coordination and a distributed enforcement approach using local policy evaluators with escalation to human oversight for cross-boundary actions. revision: yes

  3. Referee: [cross-device action placement] Section on cross-device action placement: The description of routing distributed actions based on hardware affordances and user context lacks any formal topology representation, decision procedure, or handling for dynamic device clusters, which is required to support the decoupling of intent from actuation.

    Authors: We acknowledge that the cross-device section describes routing conceptually without formal representations or procedures. As a reference architecture, the manuscript prioritizes principles over implementation details, but this leaves the decoupling claim under-specified. We will revise the section to include a graph-based topology representation and a decision procedure modeled as multi-objective optimization over affordance and context attributes, with explicit handling for dynamic clusters via discovery protocols and re-evaluation triggers. revision: yes

Circularity Check

0 steps flagged

No significant circularity in reference architecture description

full rationale

The manuscript is an engineering-oriented reference architecture proposal with no equations, derivations, fitted parameters, or mathematical claims. The three core contributions and the modeling of two coupled topologies are presented as design choices and descriptive mechanisms rather than results derived from prior steps or self-citations. No load-bearing argument reduces to a self-defined quantity or unverified self-citation chain; the central sufficiency assumption is stated explicitly as a modeling premise without circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that physical and social topologies can be coupled to support accountable execution. The Digital Twins concept is introduced as a modeling device without external validation in the abstract. No free parameters are mentioned.

axioms (1)
  • domain assumption Agents can be treated as socially situated Digital Twins that coordinate while preserving provenance, role-aware permissions, and human accountability.
    Invoked directly in the description of cross-user identity attribution in the abstract.
invented entities (1)
  • Digital Twins no independent evidence
    purpose: To enable coordination in multi-user spaces while preserving provenance and human accountability.
    Postulated in contribution (2) as the mechanism for cross-user identity attribution.

pith-pipeline@v0.9.0 · 5831 in / 1559 out tokens · 87000 ms · 2026-05-20T17:12:55.622375+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages · 6 internal anchors

  1. [1]

    Chain-of-thought prompting elicits reasoning in large language models.NeurIPS, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models.NeurIPS, 2022. 9

    work page 2022

  2. [2]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Idan Shafran, Karthik Narasimhan, and Yunhua Cao. React: Synergizing reasoning and acting in language models. InICLR, 2023

    work page 2023

  3. [3]

    Autogpt.https://github.com/significant-gravitas/AutoGPT, 2023

    Significant Gravitas. Autogpt.https://github.com/significant-gravitas/AutoGPT, 2023

    work page 2023

  4. [4]

    Langchain.https://github.com/langchain-ai/langchain, 2022

    Harrison Chase. Langchain.https://github.com/langchain-ai/langchain, 2022

    work page 2022

  5. [5]

    OpenAgents: An open platform for language agents in the wild.arXiv preprint arXiv:2310.10634, 2023

    Tianbao Xie, Fan Zhou, Zhoujun Cheng, Peng Shi, Luoxuan Weng, Yitao Liu, Toh Jing Hua, Junning Zhao, Qian Liu, Che Liu, Leo Z. Liu, Yiheng Xu, Hongjin Su, Dongchan Shin, Caiming Xiong, and Tao Yu. Openagents: An open platform for language agents in the wild. arXiv preprint arXiv:2310.10634, 2023

    work page arXiv 2023

  6. [6]

    AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation

    Qingyun Wu et al. Autogen: Enabling next-gen llm applications via multi-agent conversation. arXiv preprint arXiv:2308.08155, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  7. [7]

    Aios:Llmagentoperatingsystem

    Kai Mei et al. Aios: Llm agent operating system. arXiv preprint arXiv:2403.16971, 2024

    work page arXiv 2024

  8. [8]

    Integrating artificial intelligence into operating systems: A survey on techniques, applications, and future directions

    Xinkui Zhao, Yifan Zhang, Ziying Li, Guanjie Cheng, Jianwei Yin, Zhiquan Liu, Lufei Zhang, and Zuoning Chen. Integrating artificial intelligence into operating systems: A survey on techniques, applications, and future directions. arXiv preprint arXiv:2407.14567, 2024

    work page arXiv 2024

  9. [9]

    Openclaw: Personal ai assistant.https://github.com/Openclaw/Openclaw, 2026

    OpenClaw Contributors. Openclaw: Personal ai assistant.https://github.com/Openclaw/Openclaw, 2026

    work page 2026

  10. [10]

    Hermes agent.https://github.com/NousResearch/hermes-agent, 2026

    Nous Research. Hermes agent.https://github.com/NousResearch/hermes-agent, 2026

    work page 2026

  11. [11]

    GenericAgent: A Token-Efficient Self-Evolving LLM Agent via Contextual Information Density Maximization (V1.0)

    Jiaqing Liang, Jinyi Han, Weijia Li, Xinyi Wang, Zhoujia Zhang, et al. Genericagent: A token-efficient self-evolving llm agent via contextual information density maximization. arXiv preprint arXiv:2604.17091, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  12. [12]

    AutoDroid: LLM-powered task automation in Android

    Hao Wen, Yuanchun Li, Guohong Liu, Shanhui Zhao, Tao Yu, Toby Jia-Jun Li, Shiqi Jiang, Yunhao Liu, Yaqin Zhang, and Yunxin Liu. AutoDroid: LLM-powered task automation in Android. InProceedings of the 30th Annual International Conference on Mobile Computing and Networking, MobiCom ’24, 2024

    work page 2024

  13. [13]

    Mobile-Agent-v3: Fundamental Agents for GUI Automation

    Jiabo Ye, Xi Zhang, Haiyang Xu, Haowei Liu, Junyang Wang, Zhaoqing Zhu, Ziwei Zheng, Feiyu Gao, Junjie Cao, Zhengxi Lu, Jitong Liao, Qi Zheng, Fei Huang, Jingren Zhou, and Ming Yan. Mobile-Agent-v3: Fundamental agents for GUI automation. arXiv preprint arXiv:2508.15144, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [14]

    OS agents: A survey on MLLM-based agents for computer, phone, and browser use

    XueyuHu,TaoXiong,BiaoYi,ZishuWei,RuixuanXiao,YurunChen,JiashengYe,MeilingTao,Xiangxin Zhou, Ziyu Zhao, Yuhuai Li, Shengze Xu, Shenzhi Wang, Xinchen Xu, Shuofei Qiao, Zhaokai Wang, Kun Kuang, Tieyong Zeng, Liang Wang, Jiwei Li, Yuchen Eleanor Jiang, Wangchunshu Zhou, Guoyin Wang, Keting Yin, Zhou Zhao, Hongxia Yang, Fan Wu, Shengyu Zhang, and Fei Wu. OS ag...

    work page 2025

  15. [15]

    Generative Agents: Interactive Simulacra of Human Behavior

    Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  16. [16]

    MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Shanshan Yao, Weiyi Xu, Zhoujun Chen, Fangyin Wang, Kuncheng Xiao, Cheng Li, Qingying Wu, Zhiyuan Fan, Xiang Li, Xiaolong Wang, Jai Li, Bing Qin, Juanzi Chen, Chengming Wang, Jiannan Liu, Zhaoxiang Xiao, Peilin Zhou, Jian Gu, Jürgen Schmidhuber, and ...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  17. [17]

    JeromeH.SaltzerandMichaelD.Schroeder.Theprotectionofinformationincomputersystems.Proceedings of the IEEE, 63(9):1278–1308, 1975

    work page 1975

  18. [18]

    Zero trust architecture

    National Institute of Standards and Technology. Zero trust architecture. Special Publication 800-207, NIST, 2020. 10

    work page 2020

  19. [19]

    Owasp top 10 for large language model applications

    OWASP Foundation. Owasp top 10 for large language model applications. https://owasp.org/ www-project-top-10-for-large-language-model-applications/, 2025

    work page 2025

  20. [20]

    MemGPT: Towards LLMs as Operating Systems

    CharlesPacker,SarahWooders,KevinLin,VivianFang,ShishirG.Patil,IonStoica,andJosephE.Gonzalez. MemGPT: Towards LLMs as operating systems. arXiv preprint arXiv:2310.08560, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  21. [21]

    Digital twin in manufacturing: A categorical literature review and classification.IFAC-PapersOnLine, 51(11):1016–1022, 2018

    Werner Kritzinger, Matthias Karner, Georg Traar, Julia Henjes, and Wilfried Sihn. Digital twin in manufacturing: A categorical literature review and classification.IFAC-PapersOnLine, 51(11):1016–1022, 2018

    work page 2018

  22. [22]

    Introducing the model context protocol

    Anthropic. Introducing the model context protocol. https://www.anthropic.com/news/ model-context-protocol, 2024. 11

    work page 2024